SamplerSupreme

SamplerSupreme is the flagship of the ComfyUI Extra Samplers pack, and the name is only half a joke. It's the pack author's (Clybius, of Chroma fame) everything-but-the-kitchen-sink sampler: a dozen step methods, seven noise samplers, per-step substep averaging, edge sharpening, latent normalization, and a noise-modulation section that could be its own pack. Installed, it appears as a sampler in the dropdown, and it outputs a SAMPLER you wire into a custom sampling node like SamplerCustom or SamplerCustomNoise.

Here's the honest framing: you will never use 90% of this node, and that's fine. Its value is that the defaults are actually sensible, so you can pick it, change one or two things, and get a genuinely different character of generation. Treat it as a toolbox, not a checklist.

The inputs that actually matter for a beginner:

• step_method - the ODE solver. Default euler, plus DPM 1/2/3S, Bogacki-Shampine, RK4/RKF45, reversible variants, trapezoidal, RES, and a dynamic mode that picks a solver per step based on estimated error. "Dynamic" is the one worth a spin; it's the pack's idea of adaptive quality.
• substeps - how many times each step is averaged (ReNoise-inspired). Default 2. Raising it can buy quality at a fixed step count, at real compute cost.
• edge_enhancement - sharpens the latent with an unsharp mask plus bilateral blur, so only edges pop. Default 0.05; bump it if your output looks soft. This is why the pack depends on kornia.
• centralization (0.02) and normalization (0.01) - subtle latent stats tweaks. Small values only; they change brightness and contrast, not structure.
• noise_modulation / modulation_strength / modulation_dims - the pack's flavor of "modulate noise per step" (intensity, frequency, or a spectral signum variant). Defaults are fine; this is the rabbit-hole section.

The rest - eta, s_noise, reversible_eta, perphist, noise_sampler_type, substep_method - are for when you're deliberately chasing a specific behavior. reversible_eta only does anything with the reversible step methods, and perphist (adding the previous denoised variable via perpendicular projection) is the kind of thing you'd tune only after you've seen the base output.

If you're on a flow-matching model (Flux-style), remember the pack was mostly built for the SD/DDPM world, and the community's hard-won lesson is that exotic high-order solvers and aggressively reshaped schedules can fall apart on a straight trajectory - the sampler essay in the knowledge base calls this out repeatedly. So: on SD 1.5/SDXL, have fun; on flow models, start conservative (euler-style method, low steps) and treat every knob beyond that as an experiment with a revert button.

Installing - same as the whole pack, no model downloads:

cd ComfyUI/custom_nodes
git clone https://github.com/Clybius/ComfyUI-Extra-Samplers

restart ComfyUI (or install via ComfyUI Manager, searching "ComfyUI Extra Samplers"). Dependency is kornia, handled by Manager.

Where people get burned: they crank edge_enhancement high and substeps high and then blame the pack when images get crunchy and generation crawls. Start at defaults, change exactly one knob, and keep a fixed seed while you compare. That discipline is what separates "Supreme looks great" from "Supreme is broken" - the node is just giving you a very long lever arm.